This repository contains the scripts to reproduce the figures in the main text (Fig 2- 8) and supplementary information (Fig S1-S16) of the following article:
MohammadMehdi Kafashan, Anna Jaffe, Selmaan N Chettih, Ramon Nogueira, Iñigo Arandia-Romero, Christopher D Harvey, Rubén Moreno-Bote, Jan Drugowitsch (2021). Scaling of sensory information in large neural populations shows signatures of information-limiting correlations. Nature Communications, 12:472, 1-16; doi:org/10.1038/s41467-020-20722-y.
All scripts have been tested and run under MATLAB R2020a on MacOS Catalina. They should also work on other operating systems or later MATLAB versions. They might also run on GNU Octave, the free MATLAB alternative, but this has not been tested.
The provided figure-generating scripts don't run as is. They require per-computing Fisher information moments, fit parameters, and other interim results that take up >200Gb, such that we don't provide them here. We provide all the scripts required to compute these interim results. See Usage for how to compute them.
All scripts are written in MATLAB. To run the scripts, first, download a copy of the scripts, and extract them to a folder of your choice. Furthermore, download the neural data from Figshare and place it into the data folder.
The main directory has several folders including data, figs, fits, moment_cache, reordering_cache, shared, simData and tuning_fits folders. Make sure that the MATLAB path points to the folder to which you extracted the scripts to. Add the shared folder to the MATLAB path as there are multiple utility functions in this folder that are used to perform the analysis.
In the manuscript we number mice from 1 to 6. These numbers map to the mouse identifiers used in the scripts as follows:
mouse 1: m25 (2 sessions, a-b; 10% contrast)
mouse 2: m26 (2 sessions, a-b; 10% contrast)
mouse 3: aj42 (5 sessions, a-e; 10% contrast)
mouse 4: aj43 (7 sessions, a-g; 10% contrast)
mouse 5: aj60 (4 sessions, a-d; 10% & 25% contrast)
mouse 6: aj61 (3 sessions, a-c; 10% & 25% contrast)
Each identifier is followed by a letter that indicates the session. For example, the data for the second session of mouse 3 has the identifier aj42b.
Before executing the scripts that generate the figure plots, we need to generate a bunch of interim results that these scripts rely on. These results need to be computed for all mice/sessions and pairs of orientations. Computing these can take a considerable amount of time, and the use of computing clusters (if available) is recommended.
To perform tuning fits for recorded neural data we use the singleNeurAnlys function. For instance, to perform tuning curve fits for mouse/session m25b, we need to execute the following command in the MATLAB workspace:
singleNeurAnlys('m25b');
First, we need to compute the first and second-order moments of Fisher information for different population sizes. We can do so for mouse/session m25b by
ori1id = 1; % index of the first orientation in the virtual discrimination task
ori2id = 2; % index of the second orientation in the virtual discrimination task
conid = 1; % contrast index
popsamples = 10000; % number of bootstrap samples
computeMoments('m25b', ori1id, ori2id, conid, popsamples);
In the above,ori1id, ori2id index of the first/second orientation while conid represents the contrast index (for a dataset with X different contrast conid ranges from 1 to X). The above script need to be re-run for all mice/sessions, all possible orientation pairs (by changing ori1id and ori2id), and contrasts (by changing conid).
The computeMoments function bootstraps information increase estimates, and writes the bootstrap samples to the moment_cache folder (to a file named m25b_o1-2_c1.mat for the above example). These bootstrap samples form the basis for fitting the information increase curves.
To perform the analysis for trial-shuffled data, across trials, we need to provide a value for the subsample argument of the computeMoments script:
subsample = 'shuf';
computeMoments('m25b', ori1id, ori2id, conid, popsamples, subsample);
Above command writes the bootstrap samples to a file named m25b_o1-2_c1_shuf.mat in the moment_cache folder.
To perform the analysis only for trials with high running speed we need to instead call computeMoments as follows:
subsample = 'hispd';
computeMoments('m25b', ori1id, ori2id, conid, popsamples, subsample);
The same analysis for low running speed-trials is performed by setting subsample to lospd.
Once the Fisher information moments have been computed and saved in moment_cache, we use the scalemodel_fit script to draw posterior samples for various information scaling models for a single dataset and single pair of information as follows:
scalemodel_fit('m25b_o1-2_c1.mat');
This command fits the information scaling for mouse/session m25b, orientation identifiers 1 and 2 (o1_2), and contrast identifier 1 (c1). Specific subsets of the data (e.g., trials with high running speed), or trial-shuffled data can be fit by calling the above script for the respective moment cache file (e.g., m25b_o1-2_c1_shuf.mat).
To fit information scaling models to multiple orientation pairs simultanously, we use the script scalemodel_fit_multi. The first argument to this function specifies the name of the joint analysis (which will be the used filename to store the fits), and the remaining arguments specify the moment files that the fits are performed for. For example,
scalemodel_fit_multi("m25b_dori1_c1", 'm25b_o1-2_c1','m25b_o1-8_c1','m25b_o2-3_c1','m25b_o3-4_c1','m25b_o4-5_c1','m25b_o5-6_c1','m25b_o6-7_c1','m25b_o7-8_c1')
performs the fit across 8 different orientation pairs, all separated by 45°, for mouse/session m25b.
To compute greedy optimal orderings of neurons in terms of the information we utilize computeOrderings function for a given dataset, discrimination tasks, and contrast. The function accept a boolean argument named shuf to perform the analysis for trial-shuffled data. For instance, to perform the information estimates for optimal reordering of neurons for dataset m25 and discrimination task between direction 45° and 90° we need to execute the following commands in the MATLAB workspace:
ori1id = 1; % index of the first orientation in the virtual discrimination task
ori2id = 2; % index of the second orientation in the virtual discrimination task
conid = 1; % contrast index
computeOrderings('m25b', ori1id, ori2id, conid);
To perform the same computation for trial-shuffled data we need to execute the following commands:
ori1id = 1; % index of the first orientation in the virtual discrimination task
ori2id = 2; % index of the second orientation in the virtual discrimination task
conid = 1; % contrast index
shuffle = true; % perform trial-shuffling of neural data before optimal re-ordering analysis
computeOrderings('m25b', ori1id, ori2id, conid, shuffle);
Results of the above analysis will be stored in reordering_cache. This analysis needs to be repeated for different mice/sessions, orientation pairs, and contrasts.
For reproducing the figures in the main text execute the following commands in the workspace:
plotFig2;
plotFig3;
plotFig4;
plotFig6;
plotFig7;
plotFig8;
Except for the figures that are explicitly discussed below, all other SI figures should be executed by typing plotFigS#, where # needs to be replaced by the SI figure number.
We first need to generate the simulated neural activity using a Gaussian population model, and Linear-nonlinear-Poisson (LNP) population model (Kanitscheider et al. (2015)). We utilize simPopActivity function to generate the simulated neural activity and the neural activity of each simulation is written to simData folder with the name of sim[simid] where simid is the integer representing the simulation id and ranges from 1 to 4. Simulation details for different simid numbers are expressed as follows:
-
simid = 1- Gaussian population with limited information,$$N=1000$$ ,$$T=1000$$ ,$$I_\infty=20$$ -
simid = 2- Gaussian population with unlimited information,$$N=1000$$ ,$$T=1000$$ ,$$I_\infty=\infty$$ -
simid = 3- LNP population with limited information,$$N=2500$$ ,$$T=2500$$ -
simid = 4- LNP population with unlimited information,$$N=2500$$ ,$$T=2500$$
See Sec. 4.1 and Sec 4.2 of the SI for more details on the multivariate Gaussian population model and LNP model. After storing the simulated neural activity, computeMoments is utilized to compute the Fisher information moments for the simulated data. For instance, to compute the information moments for sim1, using only N=300 neurons and T=500 trials, we utilized the following command:
N = 300; % number of neurons to consider
T = 500; % number of trials to consider
popsamples = 10000; % number of bootstrap samples
computeMoments('sim1', N, T, popsamples);
After that, we perform the information scaling estimates using scalemodel_fit function and then execute the plotFigS7 in the MATLAB workspace to generate this SI figure.
First need to execute the single neuron analysis for all sessions of mice 5 and 6 using the singleNeurAnlys('aj60a') function as follows:
subsample = 'none';
singleNeurAnlys_mouse5n6('aj60a', subsample);
After that, we execute plotFigS12 in the MATLAB workspace to generate this SI figure.
First execute the function sim_eye_mvm to simulate neural population activity, model of Kanitscheider et al. (2015), with N=1000 and T=20000 and different combination of pn and pt as follows:
sim_eye_mvm
This function will store the information increase moments to the moment_cache. We then utilize the scalemodel_fit function to the stored moment data in the moment_cache folder to estimate the information scaling for this simulated data. After that, we execute plotFigS15 in the MATLAB workspace to generate this SI figure.
First execute the function simIcov to simulate populations of different sizes M as described in Sec. 4.1 of the SI. For that we need to execute the following command in MATLAB workspace for M=500, 1000, 5000, 10000, 15000:
M = 500; % Population size of the simulated neural activity.
simIcov(M)
The covariance statistics will be written to cov_cache folder. After executing the above command for differnet M, we execute plotFigS16 in the MATLAB workspace to generate this SI figure.